feat(registry): measurement-vendor discovery on /api/registry/agents (#3613)

[bokelley]

Summary

Implements the AAO server-side half of per-metric catalog discovery — the buyer-facing complement to the AdCP schema work that landed in #3652.

The crawler now ingests each measurement agent's get_adcp_capabilities.measurement block and stores it in the existing agent_capabilities_snapshot table (new measurement_capabilities_json JSONB column). The public /api/registry/agents endpoint gains three filters:

Param	Match	Notes
metric_id=attention_units	Exact, repeatable	JSONB containment on metrics[].metric_id
accreditation=MRC	Exact, repeatable	verified_by_aao always false — vendor-asserted, AAO does not independently verify
q=attention	Substring on metric_id	v1 scope; max 64 chars; SQL wildcards rejected outright

All three imply type=measurement when present; explicit non-measurement type returns 400. When ?capabilities=true, the response folds in capabilities.measurement_capabilities next to the existing creative_capabilities and signals_capabilities.

Design choices

Three expert reviews shaped the shape:

• Buy-side (adtech-product-expert): added accreditation= as a first-class filter (was originally only metric_id + q). Kept fuzzy q as secondary filter since planner agents handle synonym retries when the closed taxonomy was rejected. Stayed on /api/registry/agents rather than splitting /measurement-vendors — every type would otherwise need its own endpoint.
• DX: auto-scope to type=measurement when filters are present, 400 only on explicit conflict. sources counts align with filtered set (sum(sources) === count). Flat sibling under capabilities.* (don't rename to protocol_capabilities — backward compat + agent buildability).
• Security: must-fix per-field caps + 256KB DB CHECK + scriptish/URI scheme rejection at write time + ESCAPE '\\' ILIKE pattern (mirrors catalog-db.ts, not the brand-db.ts variant that omits it).

What's NOT in this PR

• q= fuzzy search across description / standard_reference / accrediting_body (v1 is metric_id only — see DX note in OpenAPI description; expand once we see real query patterns).
• Cache-key collision audit on CapabilityDiscovery.cache (security L2).
• Dashboard innerHTML audit (security M1 follow-up).
• rel="noopener noreferrer" recommendation for downstream renderers (security M2).

These are tracked as follow-ups; not blockers for this PR.

Closes #3613. Closes #3614 (direct-call vs AAO-index decision matrix folded into docs/registry/index.mdx).

Test plan

• vitest run server/tests/unit/measurement-capabilities-sanitize.test.ts — 17 sanitizer tests covering happy path, all caps, scriptish/URI rejection, control-char strip, verified_by_aao always false
• Existing registry-agents and openapi tests still pass (32/32)
• tsc --noEmit clean
• Migration 461 applies cleanly on staging
• Smoke test ?metric_id= and ?accreditation= against staging crawl data
• Confirm ?capabilities=true response shape in OpenAPI docs preview